Segment-Anything-Model: Optimized for Mobile Deployment

High-quality segmentation mask generation around any object in an image with simple input prompt

Transformer based encoder-decoder where prompts specify what to segment in an image thereby allowing segmentation without the need for additional training. The image encoder generates embeddings and the lightweight decoder operates on the embeddings for point and mask based image segmentation.

This model is an implementation of Segment-Anything-Model found here.

This repository provides scripts to run Segment-Anything-Model on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

Model Type: Semantic segmentation
Model Stats:
- Model checkpoint: vit_l
- Input resolution: 720p (720x1280)
- Number of parameters (SAMDecoder): 5.11M
- Model size (SAMDecoder): 19.6 MB

Model	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Precision	Primary Compute Unit	Target Model
SAMDecoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	8.448 ms	1 - 57 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMDecoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	5.889 ms	6 - 68 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMDecoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	5.73 ms	4 - 55 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMDecoder	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	8.67 ms	11 - 11 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart1	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	228.979 ms	12 - 181 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart1	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	161.859 ms	36 - 838 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart1	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	158.396 ms	35 - 789 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart1	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	231.597 ms	43 - 43 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart2	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	781.713 ms	12 - 147 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart2	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	567.288 ms	36 - 736 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart2	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	531.384 ms	12 - 686 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart2	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	736.185 ms	33 - 33 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart3	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	779.664 ms	12 - 159 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart3	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	576.302 ms	22 - 724 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart3	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	530.988 ms	12 - 686 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart3	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	729.557 ms	33 - 33 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart4	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	770.123 ms	12 - 151 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart4	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	569.238 ms	24 - 722 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart4	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	478.143 ms	24 - 699 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart4	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	730.872 ms	33 - 33 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart5	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	772.375 ms	0 - 133 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart5	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	568.921 ms	24 - 720 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart5	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	481.0 ms	12 - 686 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart5	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	737.772 ms	33 - 33 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart6	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	768.673 ms	12 - 148 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart6	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	568.747 ms	22 - 726 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart6	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	531.699 ms	12 - 686 MB	FP16	NPU	Segment-Anything-Model.onnx
SAMEncoderPart6	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	727.465 ms	33 - 33 MB	FP16	NPU	Segment-Anything-Model.onnx

Installation

Install the package via pip:

pip install "qai-hub-models[sam]"

Configure Qualcomm® AI Hub to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Demo off target

The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.

python -m qai_hub_models.models.sam.demo

The above demo runs a reference implementation of pre-processing, model inference, and post processing.

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.sam.demo

Run model on a cloud-hosted device

In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:

Performance check on-device on a cloud-hosted device
Downloads compiled assets that can be deployed on-device for Android.
Accuracy check between PyTorch and on-device outputs.

python -m qai_hub_models.models.sam.export

Profiling Results
------------------------------------------------------------
SAMDecoder
Device                          : Samsung Galaxy S23 (13)
Runtime                         : ONNX                   
Estimated inference time (ms)   : 8.4                    
Estimated peak memory usage (MB): [1, 57]                
Total # Ops                     : 868                    
Compute Unit(s)                 : NPU (868 ops)          

------------------------------------------------------------
SAMEncoderPart1
Device                          : Samsung Galaxy S23 (13)
Runtime                         : ONNX                   
Estimated inference time (ms)   : 229.0                  
Estimated peak memory usage (MB): [12, 181]              
Total # Ops                     : 623                    
Compute Unit(s)                 : NPU (623 ops)          

------------------------------------------------------------
SAMEncoderPart2
Device                          : Samsung Galaxy S23 (13)
Runtime                         : ONNX                   
Estimated inference time (ms)   : 781.7                  
Estimated peak memory usage (MB): [12, 147]              
Total # Ops                     : 610                    
Compute Unit(s)                 : NPU (610 ops)          

------------------------------------------------------------
SAMEncoderPart3
Device                          : Samsung Galaxy S23 (13)
Runtime                         : ONNX                   
Estimated inference time (ms)   : 779.7                  
Estimated peak memory usage (MB): [12, 159]              
Total # Ops                     : 610                    
Compute Unit(s)                 : NPU (610 ops)          

------------------------------------------------------------
SAMEncoderPart4
Device                          : Samsung Galaxy S23 (13)
Runtime                         : ONNX                   
Estimated inference time (ms)   : 770.1                  
Estimated peak memory usage (MB): [12, 151]              
Total # Ops                     : 610                    
Compute Unit(s)                 : NPU (610 ops)          

------------------------------------------------------------
SAMEncoderPart5
Device                          : Samsung Galaxy S23 (13)
Runtime                         : ONNX                   
Estimated inference time (ms)   : 772.4                  
Estimated peak memory usage (MB): [0, 133]               
Total # Ops                     : 610                    
Compute Unit(s)                 : NPU (610 ops)          

------------------------------------------------------------
SAMEncoderPart6
Device                          : Samsung Galaxy S23 (13)
Runtime                         : ONNX                   
Estimated inference time (ms)   : 768.7                  
Estimated peak memory usage (MB): [12, 148]              
Total # Ops                     : 610                    
Compute Unit(s)                 : NPU (610 ops)

How does this work?

This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:

Step 1: Compile model for on-device deployment

To compile a PyTorch model for on-device deployment, we first trace the model in memory using the jit.trace and then call the submit_compile_job API.

import torch

import qai_hub as hub
from qai_hub_models.models.sam import Model

# Load the model
model = Model.from_pretrained()
decoder_model = model.decoder
encoder_splits[0]_model = model.encoder_splits[0]
encoder_splits[1]_model = model.encoder_splits[1]
encoder_splits[2]_model = model.encoder_splits[2]
encoder_splits[3]_model = model.encoder_splits[3]
encoder_splits[4]_model = model.encoder_splits[4]
encoder_splits[5]_model = model.encoder_splits[5]

# Device
device = hub.Device("Samsung Galaxy S23")

# Trace model
decoder_input_shape = decoder_model.get_input_spec()
decoder_sample_inputs = decoder_model.sample_inputs()

traced_decoder_model = torch.jit.trace(decoder_model, [torch.tensor(data[0]) for _, data in decoder_sample_inputs.items()])

# Compile model on a specific device
decoder_compile_job = hub.submit_compile_job(
    model=traced_decoder_model ,
    device=device,
    input_specs=decoder_model.get_input_spec(),
)

# Get target model to run on-device
decoder_target_model = decoder_compile_job.get_target_model()
# Trace model
encoder_splits[0]_input_shape = encoder_splits[0]_model.get_input_spec()
encoder_splits[0]_sample_inputs = encoder_splits[0]_model.sample_inputs()

traced_encoder_splits[0]_model = torch.jit.trace(encoder_splits[0]_model, [torch.tensor(data[0]) for _, data in encoder_splits[0]_sample_inputs.items()])

# Compile model on a specific device
encoder_splits[0]_compile_job = hub.submit_compile_job(
    model=traced_encoder_splits[0]_model ,
    device=device,
    input_specs=encoder_splits[0]_model.get_input_spec(),
)

# Get target model to run on-device
encoder_splits[0]_target_model = encoder_splits[0]_compile_job.get_target_model()
# Trace model
encoder_splits[1]_input_shape = encoder_splits[1]_model.get_input_spec()
encoder_splits[1]_sample_inputs = encoder_splits[1]_model.sample_inputs()

traced_encoder_splits[1]_model = torch.jit.trace(encoder_splits[1]_model, [torch.tensor(data[0]) for _, data in encoder_splits[1]_sample_inputs.items()])

# Compile model on a specific device
encoder_splits[1]_compile_job = hub.submit_compile_job(
    model=traced_encoder_splits[1]_model ,
    device=device,
    input_specs=encoder_splits[1]_model.get_input_spec(),
)

# Get target model to run on-device
encoder_splits[1]_target_model = encoder_splits[1]_compile_job.get_target_model()
# Trace model
encoder_splits[2]_input_shape = encoder_splits[2]_model.get_input_spec()
encoder_splits[2]_sample_inputs = encoder_splits[2]_model.sample_inputs()

traced_encoder_splits[2]_model = torch.jit.trace(encoder_splits[2]_model, [torch.tensor(data[0]) for _, data in encoder_splits[2]_sample_inputs.items()])

# Compile model on a specific device
encoder_splits[2]_compile_job = hub.submit_compile_job(
    model=traced_encoder_splits[2]_model ,
    device=device,
    input_specs=encoder_splits[2]_model.get_input_spec(),
)

# Get target model to run on-device
encoder_splits[2]_target_model = encoder_splits[2]_compile_job.get_target_model()
# Trace model
encoder_splits[3]_input_shape = encoder_splits[3]_model.get_input_spec()
encoder_splits[3]_sample_inputs = encoder_splits[3]_model.sample_inputs()

traced_encoder_splits[3]_model = torch.jit.trace(encoder_splits[3]_model, [torch.tensor(data[0]) for _, data in encoder_splits[3]_sample_inputs.items()])

# Compile model on a specific device
encoder_splits[3]_compile_job = hub.submit_compile_job(
    model=traced_encoder_splits[3]_model ,
    device=device,
    input_specs=encoder_splits[3]_model.get_input_spec(),
)

# Get target model to run on-device
encoder_splits[3]_target_model = encoder_splits[3]_compile_job.get_target_model()
# Trace model
encoder_splits[4]_input_shape = encoder_splits[4]_model.get_input_spec()
encoder_splits[4]_sample_inputs = encoder_splits[4]_model.sample_inputs()

traced_encoder_splits[4]_model = torch.jit.trace(encoder_splits[4]_model, [torch.tensor(data[0]) for _, data in encoder_splits[4]_sample_inputs.items()])

# Compile model on a specific device
encoder_splits[4]_compile_job = hub.submit_compile_job(
    model=traced_encoder_splits[4]_model ,
    device=device,
    input_specs=encoder_splits[4]_model.get_input_spec(),
)

# Get target model to run on-device
encoder_splits[4]_target_model = encoder_splits[4]_compile_job.get_target_model()
# Trace model
encoder_splits[5]_input_shape = encoder_splits[5]_model.get_input_spec()
encoder_splits[5]_sample_inputs = encoder_splits[5]_model.sample_inputs()

traced_encoder_splits[5]_model = torch.jit.trace(encoder_splits[5]_model, [torch.tensor(data[0]) for _, data in encoder_splits[5]_sample_inputs.items()])

# Compile model on a specific device
encoder_splits[5]_compile_job = hub.submit_compile_job(
    model=traced_encoder_splits[5]_model ,
    device=device,
    input_specs=encoder_splits[5]_model.get_input_spec(),
)

# Get target model to run on-device
encoder_splits[5]_target_model = encoder_splits[5]_compile_job.get_target_model()

Step 2: Performance profiling on cloud-hosted device

After compiling models from step 1. Models can be profiled model on-device using the target_model. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics.

decoder_profile_job = hub.submit_profile_job(
    model=decoder_target_model,
    device=device,
)
encoder_splits[0]_profile_job = hub.submit_profile_job(
    model=encoder_splits[0]_target_model,
    device=device,
)
encoder_splits[1]_profile_job = hub.submit_profile_job(
    model=encoder_splits[1]_target_model,
    device=device,
)
encoder_splits[2]_profile_job = hub.submit_profile_job(
    model=encoder_splits[2]_target_model,
    device=device,
)
encoder_splits[3]_profile_job = hub.submit_profile_job(
    model=encoder_splits[3]_target_model,
    device=device,
)
encoder_splits[4]_profile_job = hub.submit_profile_job(
    model=encoder_splits[4]_target_model,
    device=device,
)
encoder_splits[5]_profile_job = hub.submit_profile_job(
    model=encoder_splits[5]_target_model,
    device=device,
)

Step 3: Verify on-device accuracy

To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.

decoder_input_data = decoder_model.sample_inputs()
decoder_inference_job = hub.submit_inference_job(
    model=decoder_target_model,
    device=device,
    inputs=decoder_input_data,
)
decoder_inference_job.download_output_data()
encoder_splits[0]_input_data = encoder_splits[0]_model.sample_inputs()
encoder_splits[0]_inference_job = hub.submit_inference_job(
    model=encoder_splits[0]_target_model,
    device=device,
    inputs=encoder_splits[0]_input_data,
)
encoder_splits[0]_inference_job.download_output_data()
encoder_splits[1]_input_data = encoder_splits[1]_model.sample_inputs()
encoder_splits[1]_inference_job = hub.submit_inference_job(
    model=encoder_splits[1]_target_model,
    device=device,
    inputs=encoder_splits[1]_input_data,
)
encoder_splits[1]_inference_job.download_output_data()
encoder_splits[2]_input_data = encoder_splits[2]_model.sample_inputs()
encoder_splits[2]_inference_job = hub.submit_inference_job(
    model=encoder_splits[2]_target_model,
    device=device,
    inputs=encoder_splits[2]_input_data,
)
encoder_splits[2]_inference_job.download_output_data()
encoder_splits[3]_input_data = encoder_splits[3]_model.sample_inputs()
encoder_splits[3]_inference_job = hub.submit_inference_job(
    model=encoder_splits[3]_target_model,
    device=device,
    inputs=encoder_splits[3]_input_data,
)
encoder_splits[3]_inference_job.download_output_data()
encoder_splits[4]_input_data = encoder_splits[4]_model.sample_inputs()
encoder_splits[4]_inference_job = hub.submit_inference_job(
    model=encoder_splits[4]_target_model,
    device=device,
    inputs=encoder_splits[4]_input_data,
)
encoder_splits[4]_inference_job.download_output_data()
encoder_splits[5]_input_data = encoder_splits[5]_model.sample_inputs()
encoder_splits[5]_inference_job = hub.submit_inference_job(
    model=encoder_splits[5]_target_model,
    device=device,
    inputs=encoder_splits[5]_input_data,
)
encoder_splits[5]_inference_job.download_output_data()

With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.

Note: This on-device profiling and inference requires access to Qualcomm® AI Hub. Sign up for access.

Run demo on a cloud-hosted device

You can also run the demo on-device.

python -m qai_hub_models.models.sam.demo --on-device

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.sam.demo -- --on-device

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.
QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on Segment-Anything-Model's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of Segment-Anything-Model can be found here.
The license for the compiled assets for on-device deployment can be found here

References

Community

Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
For questions or feedback please reach out to us.